Meta-analysis of Artificial Intelligence for glioblastoma diagnostics using raman spectroscopic bio-analysis for extent of resection

<h3>Introduction</h3> Glioblastoma, the deadliest primary brain tumor, presents significant surgical challenges due to its aggressive nature. Achieving gross total resection, crucial for improved survival, necessitates precise differentiation between tumor and normal brain tissue. This study evaluates the efficacy of Raman spectroscopy, coupled with machine learning, in guiding such resections. <h3>Methods</h3> A systematic, PROSPERO-registered, PRISMA-guided meta-analysis examined the performance of Raman-based methods, with a focus on machine learning techniques for processing complex spectral data. Using a DerSimonian and Laird random effects model, the analysis emphasized sensitivity, specificity, and accuracy, alongside the Area Under the Receiver Operator Characteristic Curve (AUC). <h3>Results</h3> Highlighted that support vector machines paired with Principal Components Analysis (PCA) achieved exceptional sensitivity at 99.8% and an AUC indicative of superior diagnostic ability. Comparatively, PCA with Quadratic Discriminant Analysis showed the highest accuracy. Notably, neural networks alone provided an impressive predictive performance with minimal heterogeneity. <h3>Conclusion</h3> The study concludes that artificial intelligence offers a high degree of sensitivity and specificity in the Raman spectroscopy characterization of glioblastoma. An ensemble approach incorporating PCA appears effective, yet neural networks alone also show significant promise. Continued development of these models could facilitate near real-time, intraoperative tumor classification, although further validation within the surgical context is essential.